Classification System with Attribute Verification for Explainable AI
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition technologies using deep learning face challenges in accurately recognizing target objects due to improper use of characteristics, leading to misrecognition, and lack of transparency in decision-making processes, which can result in legal and practical limitations.
Innovation Solution
A classification system that includes class data acquisition, classification, and class appropriateness check means, utilizing a database with stored class attributes and reference images to evaluate the appropriateness of candidate classes and provide explainable reasoning for classification decisions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning is used to improve classification accuracy, then recognition precision improves, but misrecognition occurs due to improper use of deep learning characteristics
Solution Approach 1:
The system introduces a feedback mechanism where classification results are verified by checking whether candidate class attributes match the input data. The verification unit provides feedback by comparing extracted attributes from the input against stored class attributes, thereby correcting potential misrecognitions caused by deep learning characteristics while preserving the high accuracy benefits.
Solution Approach 2:
The system performs preliminary action by pre-storing class attributes and reference data before classification. The verification unit uses these pre-prepared attributes to check candidate classes before finalizing classification results, preventing misrecognition before it occurs rather than correcting it afterward.
2Productivity
If deep learning models are used to increase automation, then productivity improves, but transparency of decision-making processes deteriorates
Solution Approach 1:
The verification unit acts as an intermediary between the deep learning classification model and the final decision. It translates the black-box classification output into transparent, verifiable form by checking candidate attributes against stored reference data, thereby recovering decision transparency without reducing automation efficiency.
Solution Approach 2:
The system implements self-service by automatically verifying classification results through attribute matching without requiring manual intervention. The verification unit autonomously checks candidate classes and provides explanations, maintaining high productivity while ensuring decision transparency through automated attribute verification.
3Reliability
If classification systems are made more complex to handle edge cases, then reliability improves, but device complexity increases
Solution Approach 1:
The system applies segmentation by dividing the classification process into distinct functional units: a classification unit for initial processing and a verification unit for validation. This segmentation handles edge cases reliably while keeping each unit's complexity manageable and focused on specific tasks.
Solution Approach 2:
The verification unit performs partial action by selectively verifying only candidate classes rather than reprocessing all data. This approach improves reliability for edge cases without proportionally increasing overall system complexity, as verification is applied only where needed based on classification confidence.
Data Source
AI summary
A classification system configured to obtain class determination data, perform classification on the basis of the class determination data, and check appropriateness of individual candidate classes in a candidate class list obtained via classification, by storing attributes of the candidate classes and checking the candidate classes against the stored attributes. By checking the appropriateness of individual candidate classes via a previously-constructed database in which the attributes of classes are stored, and by making comparisons as to whether the attributes of the respective candidate classes match class determination data and allowing the user to explain the reason for classification, the accuracy and explainability of the classification system can be improved.